Single Image Dehazing Via Enhanced CycleGAN

Due to the influence of atmospheric light scattering, the images acquired by outdoor imaging device in haze scene will appear low definition, contrast reduction, overexposure and other visible quality degradation, which makes it difficult to handle the relevant computer vision tasks. Therefore, image dehazing has become an important research area of computer vision. However, existing dehazing methods generally require paired image datasets that include both hazy images and corresponding ground truth images, while the recovered images are easy to occur color distortion and detail loss. In this study, an end-to-end image dehazing method based on Cycle-consistent Generative Adversarial Networks (CycleGAN) is proposed. For effectively learning the mapping relationship between hazy images and clear images, we refine the transformation module of the generator by weighting optimization, which can promote the network adaptability to scale. Then in order to further improve the quality of generated images, the enhanced perceptual loss and low-frequency loss combined with image feature attributes are constructed in the overall optimization objective of the network. The experimental results show that our dehazing algorithm effectively recovers the texture information while correcting the color distortion of original CycleGAN, and the recovery effect is clear and more natural, which better reduces the influence of haze on the imaging quality.